Summary

The formatter writes most elements to a single Vec<FormatElement> except for:

• Memoized content: It's sometimes necessary to emit the same content but using two different layouts. The formatter memoizes/interns the content's IR to avoid formatting it twice. The internet content stores the elements in its own Vec
• BestFitting layout: The BestFitting IR allows multiple different layouts for the same content. It stores the IR elements of the variants in a dedicated Vec

The fact that the Printer must support multiple Vecs add complexity to its queue because the queue needs to support a stack of slices where it pops elements from the top slice but takes the next slice if the top slice is empty.

This PR adds a few #[cold] attributes signaling that it's unlikely that the top slice is empty because a) the printer takes elements from the main top-level Vec most of the time and b) most slices contain many elements, but have only one end ;)

This gives us a solid 2-3% performance improvement in the printer (I no longer see FitsQueue::pop in the top methods when profiling the large dataset)

@BurntSushi won't believe this

Test

cargo test